How Far Are We from Using Radiomics Assessment of Gliomas in Clinical Practice?

I important characteristics from an image was described for aerial photographs as early as 1955 and eventually by Haralick et al in 1973 using computable texture features (1). Radiomics is a more recent and fancy name given to this field of study in which high-throughput data are extracted and large amounts of quantitative imaging features are generated from medical images using data-characterization algorithms and computers. In a way, it can be thought of as reverse engineering of medical images—for decades it has been the diagnostic imaging unit manufacturers’ aim to create images from data acquired from human tissue, in some cases postprocessing those data to make the images “prettier” to the viewing eye such as through the use of smoothing algorithms or improving contrast-to-noise ratio while in the process altering, hiding, or potentially losing acquired information. By reversing that process, radiomics seeks not only to go back to the vast data used to create images in the first place but also to uncover the patterns of imaging phenotypes hidden within those data that could be clinically useful. In brief, image acquisition, segmentation, feature extraction, and feature selection are some of the essential steps involved in radiomics analysis. Thus far, in tumor radiomics, commonly extracted features generally fall into four categories: first-order statistics, texture features, wavelet features (features from a transformed space), and shape features. First-order statistics derive features such as mean, variance, kurtosis, skewness, and entropy that describe the histogram distribution of the entire tumor. Texture features are called second-order statistics because they capture spatial mutual dependencies of the image voxels such as homogeneity, contrast, gray-level nonuniformity, cluster tendency, and harder-to-picture features such as short run emphasis. Two methods commonly used to derive texture features include gray-level co-occurrence matrices and gray-level run-length matrices. To complicate matters, images can be mathematically transformed to then extract features in the transform space, such as wavelet transformation yielding features dependent on spatial frequencies. Finally, shape features such as volume, surface area, sphericity, compactness, and flatness may be extracted on the basis of the shape of the tumor border. The first three of these subgroups of features are also known as “agnostic” features; these are mathematically extracted quantitative descriptors and historically have not been part of the typical radiologist’s lexicon (2). On the other hand, shape features are often referred to as “semantic” features precisely because some of these have been intrinsic to the radiologist’s lexicon for years (2), although radiomics additionally includes quantification of these features with computer assistance. Because radiology training and clinical practice focus on image pattern recognition from “processed and presented” medical images rather than tasks of quantitative data analysis, radiologists are not familiar with, let alone accustomed to, using the majority of agnostic features and terminology. Hence, to think that such features will soon be incorporated into routine radiology reports may be premature. In this issue of Radiology, Bae et al (3) extracted 796 radiomic features (702 texture features, 70 shape features, and 24 apparent diffusion coefficient [ADC] histogram features) from multiparametric MRI and used machine learning to demonstrate the added value of radiomics analysis to clinical and genetic features for survival prediction in 217 patients with glioblastoma. The authors conducted random survival forest (RSF) analysis and exploited open-source packages to build the pipeline from image processing to machine learning by using separate training (163 patients) and test (54 patients) sets to validate their results. While the benefit in area under the receiver operating characteristic curve overall was somewhat incremental in this study, the results show the potential influence that radiomic features may have in the future of image and lesion analysis. Having such a plethora of numbers and types of imaging features is exhausting to even think about. Many are abstract and hard to picture. There is the concern of overfitting a model and not learning the true basis of a decision. One way to help address the problem of overfitting is first to perform feature selection, choosing the most powerful features and reducing the number of features inputted into the model. Bae et al used variable-hunting feature selection to whittle their original 796 features down to the 18 most useful ones (two first-order, nine texture, and seven shape features). In addition to feature selection, model choice is another variable that further complicates an analysis such as this. Here, the authors chose an RSF model, using standard 10fold cross validation on the training set and validated on the test set. Compared with commonly used Cox regression, RSF has two main reported advantages: (a) RSFs are free from proportional hazard assumption and are fully nonparametric (thus, the prognostic value of the RSF model is not limited, even when some features of the model are timedependent [eg, risk associated with the feature changes How Far Are We from Using Radiomics Assessment of Gliomas in Clinical Practice?